/// <inheritdoc />
public override void SetPriorsFromPosteriors(
int[] newWorkerToOldWorkerMap,
int[] newWordToOldWordMap,
ModelPosteriors modelPosteriors)
{
base.SetPriorsFromPosteriors(newWorkerToOldWorkerMap, newWordToOldWordMap, modelPosteriors);
var biasedCommunityWordsModelPosteriors = (BiasedCommunityWordsPosteriors)modelPosteriors;
var probWordPosteriors = biasedCommunityWordsModelPosteriors.ProbWordPosterior;
AssertWhenDebugging.IsTrue(this.LabelValueCount == probWordPosteriors.Length);
this.ProbWordsPrior.ObservedValue = Util.ArrayInit(this.LabelValueCount, i => Dirichlet.Symmetric(this.VocabularySize, this.ProbWordInitialCount));
for (var labIndex = 0; labIndex < probWordPosteriors.Length; labIndex++)
{
var probWordPosterior = probWordPosteriors[labIndex];
var oldPseudoCounts = probWordPosterior.PseudoCount;
var newPseudoCounts = Vector.Constant(this.VocabularySize, this.ProbWordInitialCount, oldPseudoCounts.Sparsity);
for (var wordIndex = 0; wordIndex < newWordToOldWordMap.Length; wordIndex++)
{
var oldIdx = newWordToOldWordMap[wordIndex];
if (oldIdx >= 0)
{
newPseudoCounts[wordIndex] = oldPseudoCounts[oldIdx];
}
}
this.ProbWordsPrior.ObservedValue[labIndex] = new Dirichlet(newPseudoCounts);
}
}
/// <inheritdoc />
public override ModelPosteriors InferPosteriors(
int[][] workerLabel,
int[][] workerJudgedTweetIndex,
int[][] words = null,
int[] wordCounts = null,
int[] newWorkerToOldWorkerMap = null,
int[] newWordToOldWordMap = null,
int?[] goldLabels = null,
ModelPosteriors oldPosteriors = null,
int numIterations = 20)
{
this.ObserveLabels(workerLabel, workerJudgedTweetIndex, goldLabels);
this.ObserveWords(words, wordCounts);
if (newWorkerToOldWorkerMap == null || oldPosteriors == null)
{
this.SetDefaultPriors();
}
else
{
this.SetPriorsFromPosteriors(newWorkerToOldWorkerMap, newWordToOldWordMap, oldPosteriors);
}
// Initialize messages
var discreteUniform = Discrete.Uniform(this.NumberOfCommunities);
this.WorkerCommunityInitializer.ObservedValue = Distribution <int> .Array(Util.ArrayInit(workerLabel.Length, w => Discrete.PointMass(discreteUniform.Sample(), this.NumberOfCommunities)));
var posteriors = new BiasedCommunityWordsPosteriors();
var evidences = new List <double>();
try
{
for (var it = 1; it <= numIterations; it++)
{
this.Engine.NumberOfIterations = it;
posteriors.TrueLabel = this.Engine.Infer <Discrete[]>(this.TrueLabel);
posteriors.CommunityCpt = this.Engine.Infer <Dirichlet[][]>(this.ProbWorkerLabel);
posteriors.WorkerCommunities = this.Engine.Infer <Discrete[]>(this.Community);
posteriors.BackgroundLabelProb = this.Engine.Infer <Dirichlet>(this.ProbLabel);
posteriors.ProbWordPosterior = this.Engine.Infer <Dirichlet[]>(this.ProbWords);
if (this.HasEvidence)
{
posteriors.Evidence = this.Engine.Infer <Bernoulli>(this.Evidence);
Console.WriteLine($"Iteration {it} log evidence:\t{posteriors.Evidence.LogOdds:0.0000}");
evidences.Add(posteriors.Evidence.LogOdds);
if (ModelBase.HasConverged(evidences))
{
break;
}
}
}
}
catch (Exception e)
{
Console.WriteLine(e);
}
return(posteriors);
}
/// <inheritdoc />
public override ModelPosteriors InferPosteriors(
int[][] workerLabel,
int[][] workerJudgedTweetIndex,
int[][] words = null,
int[] wordCounts = null,
int[] newWorkerToOldWorkerMap = null,
int[] newWordToOldWordMap = null,
int?[] goldLabels = null,
ModelPosteriors oldPosteriors = null,
int numIterations = 20)
{
this.ObserveLabels(workerLabel, workerJudgedTweetIndex, goldLabels);
if (newWorkerToOldWorkerMap == null || oldPosteriors == null)
{
this.SetDefaultPriors();
}
else
{
this.SetPriorsFromPosteriors(newWorkerToOldWorkerMap, newWordToOldWordMap, oldPosteriors);
}
var posteriors = new HonestWorkerModelPosteriors();
var evidences = new List <double>();
for (var it = 1; it <= numIterations; it++)
{
this.Engine.NumberOfIterations = it;
posteriors.TrueLabel = this.Engine.Infer <Discrete[]>(this.TrueLabel);
posteriors.BackgroundLabelProb = this.Engine.Infer <Dirichlet>(this.ProbLabel);
posteriors.RandomGuessProbability = this.Engine.Infer <Dirichlet>(this.RandomGuessProbability);
posteriors.WorkerAbility = this.Engine.Infer <Beta[]>(this.Ability);
if (this.HasEvidence)
{
posteriors.Evidence = this.Engine.Infer <Bernoulli>(this.Evidence);
Console.WriteLine($"Iteration {it} log evidence:\t{posteriors.Evidence.LogOdds:0.0000}");
evidences.Add(posteriors.Evidence.LogOdds);
if (ModelBase.HasConverged(evidences))
{
break;
}
}
}
return(posteriors);
}
/// <inheritdoc />
public override void SetPriorsFromPosteriors(
int[] newWorkerToOldWorkerMap,
int[] newWordToOldWordMap,
ModelPosteriors modelPosteriors)
{
base.SetPriorsFromPosteriors(newWorkerToOldWorkerMap, newWordToOldWordMap, modelPosteriors);
var biasedWorkerModelPosteriors = (BiasedWorkerModelPosteriors)modelPosteriors;
this.ProbWorkerLabelPrior.ObservedValue = Util.ArrayInit(this.WorkerCount, input => GetCptPrior(InitialOnDiagonalPseudoCount, InitialOffDiagonalPseudoCount, this.LabelValueCount));
for (var i = 0; i < newWorkerToOldWorkerMap.Length; i++)
{
var oldIdx = newWorkerToOldWorkerMap[i];
if (oldIdx >= 0)
{
this.ProbWorkerLabelPrior.ObservedValue[i] = biasedWorkerModelPosteriors.WorkerCpt[oldIdx];
}
}
}
/// <inheritdoc />
public override void SetPriorsFromPosteriors(
int[] newWorkerToOldWorkerMap,
int[] newWordToOldWordMap,
ModelPosteriors modelPosteriors)
{
base.SetPriorsFromPosteriors(newWorkerToOldWorkerMap, newWordToOldWordMap, modelPosteriors);
var honestWorkerModelPosteriors = (HonestWorkerModelPosteriors)modelPosteriors;
this.RandomGuessPrior.ObservedValue = honestWorkerModelPosteriors.RandomGuessProbability;
this.AbilityPrior.ObservedValue = Util.ArrayInit(this.WorkerCount, input => new Beta(2, 1));
for (var i = 0; i < newWorkerToOldWorkerMap.Length; i++)
{
var oldIdx = newWorkerToOldWorkerMap[i];
if (oldIdx >= 0)
{
this.AbilityPrior.ObservedValue[i] = honestWorkerModelPosteriors.WorkerAbility[oldIdx];
}
}
}
